arXiv subjects

Hidde Lycklama

Publications and source records attributed to Hidde Lycklama.

8 recordsLinked to original sources

arXiv2026

Atlas: Efficient Verifiable Semantic Search

Semantic search is a core primitive of modern applications, powering recommender systems, web search, and retrieval-augmented generation for language models. The provider controls the index and query execution, leaving clients to trust that results come from the right algorithm over the intended index. A provider may truncate search to cut cost, bias results, or otherwise deviate from the specified execution undetected. Verifiability can remove this trust assumption by proving that results follow the agreed algorithm over a committed index. Realizing this efficiently is hard, as retrieval at scale relies on HNSW, a graph-based algorithm whose data-dependent traversal maps poorly onto the fixed constraint systems of zero-knowledge proofs. Prior verifiable systems therefore target regular, cluster-based indices that are easier to encode, sacrificing the recall of graph-based search. We present Atlas, a system that lets a provider prove a query was answered correctly against its committed index without revealing the index. At its core is a new zero-knowledge proof for HNSW search, built on three techniques: preprocessing that shifts all database-dependent cost offline, so per-query proving scales with the traversal rather than the database; a restructuring of HNSW into a fixed-size-state procedure that we prove returns the same result; and a timestep-tagged batching that merges the per-step arguments of the entire traversal into one. Atlas is the first to demonstrate verifiable graph-based search at scale, proving a query in under a second on the SIFT1M benchmark and in 2.0 seconds at 100 million vectors, while maintaining the recall of plaintext HNSW and revealing nothing about the index beyond the result. In a complete RAG pipeline, Atlas' proven retrieval preserves end-to-end answer quality, and reaches higher quality at lower proving cost than all prior verifiable retrieval systems.

cs.CR↗

arXiv2025

UTrace: Poisoning Forensics for Private Collaborative Learning

Privacy-preserving machine learning (PPML) systems enable multiple data owners to collaboratively train models without revealing their raw, sensitive data by leveraging cryptographic protocols such as secure multi-party computation (MPC). While PPML offers strong privacy guarantees, it also introduces new attack surfaces: malicious data owners can inject poisoned data into the training process without being detected, thus undermining the integrity of the learned model. Although recent defenses, such as private input validation within MPC, can mitigate some specific poisoning strategies, they remain insufficient, particularly in preventing stealthy or distributed attacks. As the robustness of PPML remains an open challenge, strengthening trust in these systems increasingly necessitates post-hoc auditing mechanisms that instill accountability. In this paper we present UTrace, a framework for user-level traceback in PPML that attributes integrity failures to responsible data owners without compromising the privacy guarantees of MPC. UTrace encapsulates two mechanisms: a gradient similarity method that identifies suspicious update patterns linked to poisoning, and a user-level unlearning technique that quantifies each user's marginal influence on model behavior. Together, these methods allow UTrace to attribute model misbehavior to specific users with high precision. We implement UTrace within an MPC-compatible training and auditing pipeline and evaluate its effectiveness on four datasets spanning vision, text, and malware. Across ten canonical poisoning attacks, UTrace consistently achieves high detection accuracy with low false positive rates.

cs.CR↗

arXiv2025

Artemis: Efficient Commit-and-Prove SNARKs for zkML

Ensuring that AI models are both verifiable and privacy-preserving is important for trust, accountability, and compliance. To address these concerns, recent research has focused on developing zero-knowledge machine learning (zkML) techniques that enable the verification of various aspects of ML models without revealing sensitive information. However, while recent zkML advances have made significant improvements to the efficiency of proving ML computations, they have largely overlooked the costly consistency checks on committed model parameters and input data, which have become a dominant performance bottleneck. To address this gap, this paper introduces a new Commit-and-Prove SNARK (CP-SNARK) construction, Artemis, that effectively addresses the emerging challenge of commitment verification in zkML pipelines. In contrast to existing approaches, Artemis is compatible with any homomorphic polynomial commitment, including those without trusted setup. We present the first implementation of this CP-SNARK, evaluate its performance on a diverse set of ML models, and show substantial improvements over existing methods, achieving significant reductions in prover costs and maintaining efficiency even for large-scale models. For example, for the VGG model, we reduce the overhead associated with commitment checks from 11.5x to 1.1x. Our results indicate that Artemis provides a concrete step toward practical deployment of zkML, particularly in settings involving large-scale or complex models.

cs.CR↗

arXiv2025

DPolicy: Managing Privacy Risks Across Multiple Releases with Differential Privacy

Differential Privacy (DP) has emerged as a robust framework for privacy-preserving data releases and has been successfully applied in high-profile cases, such as the 2020 US Census. However, in organizational settings, the use of DP remains largely confined to isolated data releases. This approach restricts the potential of DP to serve as a framework for comprehensive privacy risk management at an organizational level. Although one might expect that the cumulative privacy risk of isolated releases could be assessed using DP's compositional property, in practice, individual DP guarantees are frequently tailored to specific releases, making it difficult to reason about their interaction or combined impact. At the same time, less tailored DP guarantees, which compose more easily, also offer only limited insight because they lead to excessively large privacy budgets that convey limited meaning. To address these limitations, we present DPolicy, a system designed to manage cumulative privacy risks across multiple data releases using DP. Unlike traditional approaches that treat each release in isolation or rely on a single (global) DP guarantee, our system employs a flexible framework that considers multiple DP guarantees simultaneously, reflecting the diverse contexts and scopes typical of real-world DP deployments. DPolicy introduces a high-level policy language to formalize privacy guarantees, making traditionally implicit assumptions on scopes and contexts explicit. By deriving the DP guarantees required to enforce complex privacy semantics from these high-level policies, DPolicy enables fine-grained privacy risk management on an organizational scale. We implement and evaluate DPolicy, demonstrating how it mitigates privacy risks that can emerge without comprehensive, organization-wide privacy risk management.

cs.CR↗

arXiv2024

Fragile Giants: Understanding the Susceptibility of Models to Subpopulation Attacks

As machine learning models become increasingly complex, concerns about their robustness and trustworthiness have become more pressing. A critical vulnerability of these models is data poisoning attacks, where adversaries deliberately alter training data to degrade model performance. One particularly stealthy form of these attacks is subpopulation poisoning, which targets distinct subgroups within a dataset while leaving overall performance largely intact. The ability of these attacks to generalize within subpopulations poses a significant risk in real-world settings, as they can be exploited to harm marginalized or underrepresented groups within the dataset. In this work, we investigate how model complexity influences susceptibility to subpopulation poisoning attacks. We introduce a theoretical framework that explains how overparameterized models, due to their large capacity, can inadvertently memorize and misclassify targeted subpopulations. To validate our theory, we conduct extensive experiments on large-scale image and text datasets using popular model architectures. Our results show a clear trend: models with more parameters are significantly more vulnerable to subpopulation poisoning. Moreover, we find that attacks on smaller, human-interpretable subgroups often go undetected by these models. These results highlight the need to develop defenses that specifically address subpopulation vulnerabilities.

cs.LG↗

arXiv2024

Holding Secrets Accountable: Auditing Privacy-Preserving Machine Learning

Recent advancements in privacy-preserving machine learning are paving the way to extend the benefits of ML to highly sensitive data that, until now, have been hard to utilize due to privacy concerns and regulatory constraints. Simultaneously, there is a growing emphasis on enhancing the transparency and accountability of machine learning, including the ability to audit ML deployments. While ML auditing and PPML have both been the subjects of intensive research, they have predominately been examined in isolation. However, their combination is becoming increasingly important. In this work, we introduce Arc, an MPC framework for auditing privacy-preserving machine learning. At the core of our framework is a new protocol for efficiently verifying MPC inputs against succinct commitments at scale. We evaluate the performance of our framework when instantiated with our consistency protocol and compare it to hashing-based and homomorphic-commitment-based approaches, demonstrating that it is up to 10^4x faster and up to 10^6x more concise.

cs.CR↗

arXiv2023

Cohere: Managing Differential Privacy in Large Scale Systems

The need for a privacy management layer in today's systems started to manifest with the emergence of new systems for privacy-preserving analytics and privacy compliance. As a result, many independent efforts have emerged that try to provide system support for privacy. Recently, the scope of privacy solutions used in systems has expanded to encompass more complex techniques such as Differential Privacy (DP). The use of these solutions in large-scale systems imposes new challenges and requirements. Careful planning and coordination are necessary to ensure that privacy guarantees are maintained across a wide range of heterogeneous applications and data systems. This requires new solutions for managing and allocating scarce and non-replenishable privacy resources. In this paper, we introduce Cohere, a new system that simplifies the use of DP in large-scale systems. Cohere implements a unified interface that allows heterogeneous applications to operate on a unified view of users' data. In this work, we further address two pressing system challenges that arise in the context of real-world deployments: ensuring the continuity of privacy-based applications (i.e., preventing privacy budget depletion) and effectively allocating scarce shared privacy resources (i.e., budget) under complex preferences. Our experiments show that Cohere achieves a 6.4--28x improvement in utility compared to the state-of-the-art across a range of complex workloads.

cs.CR↗

arXiv2023

RoFL: Robustness of Secure Federated Learning

Even though recent years have seen many attacks exposing severe vulnerabilities in Federated Learning (FL), a holistic understanding of what enables these attacks and how they can be mitigated effectively is still lacking. In this work, we demystify the inner workings of existing (targeted) attacks. We provide new insights into why these attacks are possible and why a definitive solution to FL robustness is challenging. We show that the need for ML algorithms to memorize tail data has significant implications for FL integrity. This phenomenon has largely been studied in the context of privacy; our analysis sheds light on its implications for ML integrity. We show that certain classes of severe attacks can be mitigated effectively by enforcing constraints such as norm bounds on clients' updates. We investigate how to efficiently incorporate these constraints into secure FL protocols in the single-server setting. Based on this, we propose RoFL, a new secure FL system that extends secure aggregation with privacy-preserving input validation. Specifically, RoFL can enforce constraints such as $L_2$ and $L_\infty$ bounds on high-dimensional encrypted model updates.

cs.CR↗